Abstract:
Ride-hailing applications predict the fare for each trip at the beginning of the ride.
Accurate prediction fare prediction helps improve the user experience and trip
conversion. We aim to build a machine learning model that classify the ride fare
estimation. This model helps improve the fare estimate model and feature analysis
suggest the key impact factors in ride fare. A selected sample of trips, the fare
predictions are labeled as either correct or incorrect. This project aims to develop an
accurate and explainable machine learning model to classify these fare predictions. Key
features were manually engineered from the provided dataset, including calculated
distances, driving times, and various fare-related ratios, to enhance model performance.
The dataset was carefully preprocessed to handle missing values and imbalances in the
label distribution. Various machine learning models, including Random Forest,
XGBoost, and LightGBM, were employed, with XGBoost proving to be the most
effective individual model. To further enhance accuracy, a filtering ensemble technique
was used to combine these models, achieving an impressive 98.32% accuracy. The
explainability of the model was analyzed using SHAP values, which revealed that the
most significant factors influencing predictions were ride rate, ride distance, and waiting
time. Additionally, other factors such as ride hour, ride day, and ride month also played
a role in the model’s performance. This approach provides a robust and interpretable
solution to the problem of ride fare classification. The study highlights that fare
misclassification is analogous to an outlier detection problem, where incorrect fares
often exhibit extreme rates in terms of fare per distance or driving time. The findings
suggest that further improvements could be made by incorporating additional features
related to ride penalties.
